When someone asks a generic AI chatbot about their health condition, two things happen simultaneously. First, the AI generates a fluent, confident-sounding response. Second, there is a meaningful probability that parts of that response are factually wrong.
This is not a hypothetical concern. A 2024 study in JAMA Network Open found that large language models answering common patient health questions produced accuracy rates between 60% and 85% depending on the topic. The errors were not random — they clustered around drug interactions, dosing information, and rare conditions, precisely the areas where incorrect information is most dangerous.
The problem is architectural. General-purpose AI models like GPT-4 or Claude are trained on enormous text corpora that include medical information, but also include outdated textbooks, contradictory sources, patient forum speculation, and marketing material. The model has no structural understanding of which sources are reliable. It generates text that is statistically likely to follow the input — not text that is medically verified.
Knowledge graphs offer a fundamentally different approach.
What Is a Knowledge Graph?
A knowledge graph is a structured database that represents real-world entities and the relationships between them. Unlike a document collection, which stores information as blocks of text, a knowledge graph stores information as explicit, typed connections: Disease A is treated by Drug B. Gene C is associated with Disease D. Drug E interacts with Drug F.
This structure matters because it makes relationships queryable, verifiable, and traceable. When a system answers "What drugs treat condition X?", a knowledge graph can return a specific, sourced answer rather than generating one from statistical patterns in text.
In healthcare, knowledge graphs represent medical knowledge the way a clinician's mental model works: as a network of interconnected facts, not as a collection of articles to summarize.
PrimeKG: The Knowledge Graph Behind PatientSupport.AI
PrimeKG (Precision Medicine Knowledge Graph) was developed by researchers at Harvard Medical School and published in Nature Scientific Data in 2023 (Chandak et al.). It is one of the most comprehensive biomedical knowledge graphs available for research use.
PrimeKG integrates data from 20 high-quality biomedical resources into a unified graph containing:
- 17,080 diseases with clinical descriptions and relationships
- 29,786 genes mapped to disease associations
- 7,957 drugs with indication, contraindication, and interaction data
- 4,050,249 relationships connecting diseases, genes, drugs, biological processes, and anatomical structures
- 10 relationship types including drug-disease treatment, gene-disease association, drug-drug interaction, and disease-disease comorbidity
How Knowledge Graph Grounding Works
When PatientSupport.AI receives a question about a health condition, the process differs fundamentally from a generic chatbot:
Generic chatbot approach: 1. Receive question: "What conditions are related to Type 2 diabetes?" 2. Search training data (internet text) for statistical patterns 3. Generate a response based on what tokens are likely to follow 4. No mechanism to verify whether the output is factually correct
Knowledge-graph-grounded approach: 1. Receive question: "What conditions are related to Type 2 diabetes?" 2. Query PrimeKG for all diseases with comorbidity relationships to Type 2 diabetes 3. Retrieve structured, sourced connections (cardiovascular disease, diabetic neuropathy, diabetic retinopathy, chronic kidney disease, etc.) 4. Use the language model (Groq Llama 70B) to present this structured data in natural language 5. The facts come from the graph; the language model provides the presentation
This distinction is critical. In the grounded approach, the language model is not inventing medical facts — it is translating verified structured data into readable text. The knowledge graph provides the what; the language model provides the how to say it.
Why This Reduces Hallucinations
Hallucination in AI refers to the generation of plausible-sounding but factually incorrect information. In healthcare contexts, hallucination can produce:
- Invented drug interactions that do not exist
- Incorrect dosing information
- Fabricated clinical trial results
- Conditions described with wrong symptoms
- Relationships between diseases that have no medical basis
1. Constrained output space. The model's responses are anchored to verified facts in the graph, not generated from open-ended text patterns. 2. Traceable claims. Each relationship in PrimeKG comes from a specific source database that can be cited and verified. 3. Structured relationships. "Drug A treats Disease B" is either in the graph or it is not. There is no ambiguity about what the system knows versus what it is guessing. 4. Coverage awareness. A well-designed grounded system can identify when a question falls outside the knowledge graph's coverage and say "I don't have reliable data on this" rather than fabricating an answer.
Important Caveat
Knowledge graph grounding significantly improves accuracy, but it does not make AI health tools infallible. Limitations include:
- Knowledge graph completeness. PrimeKG covers 17,000+ diseases but not every medical fact. Emerging conditions, rare case presentations, and evolving treatment protocols may not be represented.
- Temporal lag. Knowledge graphs are updated periodically, not in real-time. The latest clinical trial results or drug approvals may not be reflected immediately.
- Language model errors. Even when grounded in correct data, the language model can still misinterpret or misrepresent the data it is given. The translation from structured data to natural language is not perfect.
- Contextual limitations. A knowledge graph knows that Drug A treats Disease B, but it does not know your specific medical history, allergies, current medications, or lab values.
Comparing Approaches: Generic vs. Grounded
| Feature | Generic AI Chatbot | Knowledge-Graph-Grounded AI | |---------|-------------------|----------------------------| | Data source | Internet text corpus | Curated biomedical databases | | Fact verification | None (statistical generation) | Sourced from graph relationships | | Hallucination rate | 15-40% for health topics | Significantly reduced (not zero) | | Coverage | Broad but unreliable | Narrower but verified | | Drug interactions | May invent interactions | Returns only documented interactions | | Rare diseases | Often inaccurate | Covers 17,000+ diseases with structured data | | Source citation | Typically cannot cite sources | Can trace claims to source databases | | Update mechanism | Full retraining (expensive) | Graph update (incremental) |
The Broader Landscape of Health Knowledge Graphs
PrimeKG is not the only biomedical knowledge graph. Others in the research and clinical space include:
- UMLS (Unified Medical Language System) — maintained by the National Library of Medicine, the foundational medical terminology and relationship resource
- SNOMED CT — the comprehensive clinical terminology used in electronic health records worldwide
- DrugBank — detailed drug data including interactions, targets, and pharmacology
- Disease Ontology — standardized disease classification with cross-references to ICD, MeSH, and other systems
Why This Matters for Patient Support
The connection between knowledge graphs and patient support may not be obvious, but it is direct:
1. Patients with chronic conditions need reliable information. The internet is full of health information of wildly varying quality. A system grounded in curated medical databases provides more reliable starting points for patient education.
2. Comorbidity understanding. Most chronic disease patients have more than one condition. Knowledge graphs explicitly model how conditions relate — which helps patients understand why their treatment plan addresses multiple issues.
3. Preparation for clinical conversations. Patients who understand their condition's relationships — to other diseases, to medications, to biological processes — can have more productive conversations with their healthcare providers. See: How to Talk to Your Doctor After Using an Online Support Resource.
4. Complementing peer support. Support groups provide emotional connection and lived experience. Knowledge-graph-grounded tools provide structured medical context. Together, they address both the emotional and informational needs of patients.
How to Use PatientSupport.AI
PatientSupport.AI is free to use without creating an account. You can optionally create a free account to save your conversation history. The tool uses PrimeKG (Harvard, Nature Scientific Data 2023) and Groq Llama 70B to provide health information grounded in peer-reviewed biomedical data.
To get the most value:
- Ask specific questions about your condition, its comorbidities, or disease mechanisms
- Use the information as a starting point for conversations with your healthcare provider
- Cross-reference any information with your medical team before making treatment decisions
- Remember that AI tools — even grounded ones — can make errors
PatientSupport.AI is a health information tool, not a diagnostic or prescriptive service. It is not a replacement for professional medical care. Always consult your healthcare provider for medical decisions.